基于原型与正交投影学习的图像集分类算法

为了利用图像集中的集合信息来提高图像识别精度以及对图像变化的鲁棒性，从而大幅降低诸如姿态、光照、遮挡和未对齐等因素对识别精度的影响，提出了一种用于图像集分类的图像集原型与投影学习算法（LPSOP）。该算法针对每个图像集学习有代表性的点（原型）以及一个正交的全局投影矩阵，使得在目标子空间的每个图像集可以被最优地分类到同类的最近原型集中。用学习到的原型来代表该图像集，既能降低冗余图像干扰，又能减少存储和计算开销，学习到的投影矩阵则能够大幅提高分类精度与噪声鲁棒性。在UCSD/Honda、CMU MoBo和YouTube celebrities这三个数据集上的实验结果表明，LPSOP比目前流行的图像集分类算法具有更高的识别精度和更好的鲁棒性。

Learning of prototype set and orthogonal projection for image set classification

In order to improve the identification accuracy and the robustness by using collection information of the image set, and hence greatly reduce the influence of posture, light, misalignment and so on, this paper developed a novel method, called learning prototype set and orthogonal projection for image set classification(LPSOP), which simultaneously learnt the representatives(prototypes) and a linear discriminative projection for each image set, making any image set in the target subspace can be classified into its nearest neighbor prototype optimally. In addition, the learned representatives not only reduced redundant image noise but also reduced the consumption of time and memory. At the same time, the projection matrix greatly improved the classification accuracy and noise robustness. Experimental results on UCSD/Honda, CMU MoBo and YouTube databases prove that compared to state-of-the-art learning methods, LPSOP has higher recognition accuracy and better robustness.